FashionKG-RAG tests how AI reads a fashion question

The experiment

FashionKG-RAG’s authors retrieve graph paths by matching entities, rank them against a shortened question, and check candidates against the original wording. A path links concepts through named relationships. In their trench-coat example, they remove a skirt reference and retain the wartime connection.

They extract entities and relationships from textbook chapters, retain Chinese metadata and translate triples into English. They extend the graph with DBpedia and web-assisted generation. Their checks include human review, additional models, classification and deduplication.

They test 2,153 graph-derived multiple-choice questions, plus 200 textbook-based questions that four annotators wrote without graph access.

The main retrieval-method comparison holds FashionEcoKG constant. Table VI adds LightRAG and LinearRAG, which build their own graphs from textbook chunks. With Qwen-2.5-7B answering, the authors report PG-RAG at 74.17%, no retrieval at 67.04%, LightRAG at 63.87% and LinearRAG at 64.10%.

On the separate human-written set, they report a best PG-RAG accuracy of 85.5% without identifying the generator. These are arXiv v1 preprint results.

Editorial interpretation

Consider a hypothetical archive brief: find garments that resemble a chosen silhouette, restrict references to a named decade and exclude ceremonial dress. A shortened request could help locate candidates. Have the system display the paths it selected, and check those paths and the final answer against the original date and exclusion. Include a case where a convincing answer violates one condition.

For an internal trial, write questions before inspecting the proposed graph. Keep the source collection fixed when comparing retrieval methods, and run a separate comparison for the complete packages, including the work required to build each knowledge base. That separation would help a buyer distinguish better selection from better source preparation. It would make the staff time spent preparing references part of the decision.

Editorial judgment belongs in graph construction. Record a relation’s source passage, preserve the original wording beside its translation and make room for competing accounts. A label such as “influenced by” calls for evidence of that connection. For a studio’s own evaluation, inspect the historical claim as well as the relation’s wording.

We have not replicated this experiment or tested a studio deployment. Before adoption, we would want an evaluator to record unsupported answers and the time a researcher spends correcting them.